Hierarchical Transformer for Efficient CT Image Super-Resolution

This paper proposes a super-resolution (SR) reconstruction method to improve the poor quality of low-resolution (LR) computed tomography (CT) images. Transformer-based methods have shown promising performance in image SR. However, there are two obstacles and key challenges: (1) The hierarchical information of images is not fully utilized; (2) The potential of the self-attention (SA) mechanism has not been fully exploited. To this end, we propose a Hierarchical Transformer for efficient CT image Super-Resolution(HTSR). Our HTSR can capture image hierarchical information from local, regional, and global. Specifically, to achieve efficient region modelling, we design a Content-Aware Transformer Block (CATB), which is a coarse-to-fine image block selection strategy to model long-range dependencies of semantically consistent information and eliminate interference from bad information. Moreover, we introduce Transposed Transformer Block (TTB), which can capture the long-range dependency and is applicable to large-sized images. Finally, combined with the Spatial window Transformer Block (SWTB), we design a network architecture of encoder-decoder, achieving an effective fusion of hierarchical information. Extensive experimental results have shown that the proposed method achieves good performance.